Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
AngleRoCL: Angle-Robust Concept Learning for Physically View-Invariant Adversarial Patches
Authors: Wenjun Ji, Yuxiang Fu, Luyang Ying, Deng-Ping Fan, Yuyi Wang, Ming-Ming Cheng, Ivor Tsang, Qing Guo
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive simulation and physical-world experiments on five SOTA detectors across multiple views, we demonstrate that Angle Ro CL significantly enhances the angle robustness of T2I adversarial patches compared to baseline methods. Our patches maintain high attack success rates even under challenging viewing conditions, with over 50% average relative improvement in attack effectiveness across multiple angles. This research advances the understanding of physically angle-robust patches and provides insights into the relationship between textual concepts and physical properties in T2I-generated contents. |
| Researcher Affiliation | Collaboration | Wenjun Ji1,2 Yuxiang Fu1,2 Luyang Ying1,2 Deng-Ping Fan1,2 Yuyi Wang3 Ming-Ming Cheng1,2 Ivor Tsang4 Qing Guo1,2 1NKIARI, Shenzhen Futian 2VCIP, CS, Nankai University 3CRRC Zhuzhou Institute 4CFAR and IHPC, Agency for Science, Technology and Research (A*STAR) |
| Pseudocode | No | No explicit pseudocode or algorithm blocks were found in the main body of the paper. The methodology is described textually and with mathematical formulations. |
| Open Source Code | Yes | We released our code in https://github.com/tsingqguo/anglerocl. |
| Open Datasets | Yes | Following the evaluation protocol in [58], we adopt the nu Image dataset [14], selecting one representative image from each of the six car-mounted camera views (front, front left, front right, back, back left, back right) for digital evaluation. |
| Dataset Splits | No | The paper does not explicitly provide training/test/validation dataset splits for the Angle Ro CL model. It describes selecting images from the nu Image dataset for evaluation purposes, but not traditional splits for model training. |
| Hardware Specification | Yes | Experiments ran on 2 NVIDIA 3090 GPUs. |
| Software Dependencies | Yes | We implement our approach using the Stable Diffusion v1.5 as the base diffusion model with DPMSolver++ for denoising. ... For consistency with prior work [48, 58], we employ Stable Diffusion v1.5 [46] as our image generator. Our evaluation spans multiple object detection architectures, including YOLOv5 [29], YOLOv3 [44], Faster R-CNN [44], and RT-DETR [61]. We additionally evaluate against YOLOv10 [53]... For implementation frameworks, YOLOv5 and YOLOv10 are evaluated using the API from ultralytics, while Faster R-CNN, YOLOv3, and RT-DETR are implemented through the MMDetection framework... |
| Experiment Setup | Yes | We set the classifier-free guidance scale to 7.5 and use 25 denoising steps. For the angle-robust concept, we use CLIP embedding of <angle-robust> as the initialization... During the training process, we sample 9 angles, i.e., { 72 , 54 , 36 , 18 , 0 , 18 , 36 , 54 , 72 }... The detection loss parameter y and the scaling factor λ are respectively set to 0.8 and 10. We train for 50,000 steps using Adam W optimizer with learning rate 10 4, updating only the <angle-robust> embedding while keeping all other parameters frozen. |